Forum Moderators: open
Although a human editor might say, hmm this site would best fit this category, he may be unaware of another category outside his / her field that is a better match under a different hierarchy.
That's true, but the user of your directory may face the same problem - he will look for a website in his field, while your AI-brain put it in some other place. Or will each one website be inserted to all relevant categories?
Or will each one website be inserted to all relevant categories?
We previously have and currently are considering that exact outcome. You have brought up an excellent point.
Relevancy, content, and meaning are extremely hard things to pin down from a human's perspective ;) .
Nor am I suggesting that the trained program won't be able to do more sites more quickly than ODP editors.
I'm suggesting that you cheat and use ODP listings to help train your system. THAT information -- four million listings already categorized by the public, representing several hundred man-years of work by those very same public users, is a resource you'll find hard to duplicate.
This should be a really potential development, if you can get the end results.
I have stickied you requesting the url. can i take a peek at your work?
Good luck
Cheers
once you expand beyond the sex&drugs categories
I'm not sure what you are reffering to? The current demo (as of 02-26-04) works on ALL main level topics and the hundreds of thousands of sub categories. We are beginning to think that perhaps the steps involved in bringing this to market may be too confusing to users, and perhaps we should refrain from updates.. to avoid misconceptions and miscommunication..
The categorizers current placement of "webmasterworld.com"
RESULTS FOR [webmasterworld.com:...]
News » Personalized News
The following are also suggested matches for this URL:
[MOVE HERE] News » Breaking News
[MOVE HERE] Adult » News
[MOVE HERE] Computers » Bulletin Board Systems » Web Discussion Boards
Hope tonights algo tweaks mix the results more inline ;)
I just tested their current auto categories and should say, i was quite impressed.
I tried categorizing a few sites and it did a decent work. out of 3 sites, i tried only one site wasn't correctly categorised.
It actually does a correct classification of deep level cat's but the top levels seems to be a bit not relevant. that may be b'cos they currently are testing it using dmoz cat structure.
They probably have to move out of that dmoz structure and have their own categories and/or let the script do the creation of categories.
BTW, have you activated the option of creating its own categories, if need be when the script tries to classify?
I feel, this has lots of potential and it sure is an interesting development.
Good luck
Cheers
Thank you to eveyone who has provided us with input regarding the results thus far!
[edited by: skibum at 7:50 pm (utc) on Mar. 13, 2004]
Keep tweaking :)
WBF
I am a developer working on this categorizer project. We originally started with a category structure based on a yellow pages site (about 5000 categories). After getting that to work, we used most of DMOZ's structure (500,000 categories) and got superb results using substantially the same algorithm. So our setup is completely nonreliant upon the specifics of the directory (the mathematical equations could work on ANY directory sturcuture).
If you have any code that you could put online, we'd love to see how your categorizer works!
I beleive they are big enough to be refrerred to:
vivisimo.com
PR=8
They are listed in dmoz at metasearch category.
Auto categorisation of content against pre-defined categories is nothing new.
What is new is launching a website that does spidering/categorisation on the fly with response times measured in tens of seconds not milliseconds :)
This spider/categorise on-demand approach simply will not scale with volumes of users.
The alternative is to pre-spider and categorise all the content which just leaves you with the problem of spidering 4 billion plus pages.
I challenge anyone to post or sticky me a another directory that auto-categorizes websites via a spider within typically 3 seconds
It is relevant to mention netinsert.com which indexes and categorizes a web page via a spider. In that case, the classification is not automatic but distributed to the end users. The distribution is accomplished using a meta tag which refers to a category in the directory taxonomy.
netinsert's requirement to "suggest" a website
In order to add a web page to the directory, please complete the following steps:Navigate to the subject category where the web page shall be added. Select the subject category either in the directory for World or in a country or a region that is relevant (more on selecting directory...)
Submit the URL to the web page in the top right hand text field, and press the button for "Insert Page"
Follow the instructions on the screen (a meta tag must be inserted in the html source code of the page)
*owner edit
the classification is not automatic
We looked at vivisimo et al, and were quite disappointed with the results. It appears to make categories by looking at continually repeated keywords in the search results, leading to some very odd and irrelevant categories.
Auto categorisation of content against pre-defined categories is nothing new.
Thanks, we have been looking for this kind of information. Since it's nothing new I'm sure you have a few URLs you could supply of sites that are doing the same thing?
What is new is launching a website that does spidering/categorisation on the fly with response times measured in tens of seconds not milliseconds
Maybe you have different standards but I hardly call posting a piece of beta code on a WW thread "launching a website". It's still very much under development, not optimized for speed (obviously!), and just running on an old linux box at the moment.
We're very stoked about the fact that it works so accurately in the first place and we've been spending weeks making it continually work better. Thanks to the great positive feedback and constructive criticism we've received from numerous people here, we have been able to develop something that is--as far as I can tell--entirely unique.
Can I point you at sites that do on demand-spidering and auto-categorisation? No - and the reasons are commercial NOT technical.
There are a LOT of neat tricks that can be done with search technology that never get launched to an audience the size of the web. When launching a site you have to balance the computational costs involved and the possible payback.
I'm not questioning your technology, more your intended application of it on the web.
If you want to look at existing auto-categorisation tools, one example is Verity - they have been doing it for years. They can even generate category hierarchies automatically from content.
Yes, a huge site that would accept sites on the fly for auto-categorization would require a massive amount of computing power. At this point we're really just making it work as accurately as possible, then we want to explore all the potential applications for it.
Thanks for the info on Verity, I had not heard of them previously but I will certainly check them out.
To have an idea about what other people are doing, just google for "automatic classification systems" (include the quotes in your search).
1st generation directories
Staff categorization. A small group of people, either employed or volunteers, categorizes the web. Data for a web page is manually submitted.
Pros:
- a high degree of order in the directory
Cons:
- a losing struggle to the ever growing web in terms of new submissions and updates of current listings
2nd generation directories
User categorization. Each web page owner categorizes his own web page using a global taxonomy. Data for a web page is spidered.
Pros:
- broad user involvement
- instant submission and automatic updates
- classification becomes a web page property which can be used in higher order search algorithms.
Cons:
- less degree of order in the directory.
3rd generation directories
Automatic categorization. AI like algorithms classify a web page without any human review. Data for a web page is spidered.
Pros:
- completely automated procedure
- instant submission and automatic updates
Cons:
- very difficult to define robust algorithms that can be applied generally (people have worked on AI systems since the 1960s)
- reverse engineering of algorithms (i.e. "seo" applied to categorization)
